The present invention enables service agents in a call center to monitor portions of calls that are especially error-prone (e.g., automated speech recognition of an address, etc.), while portions of calls that are not error-prone occur without any human monitoring. An interactive voice response (IVR) system script (e.g., a VXML script, etc.) is divided into a plurality of independent dialog sequences, each of which is assigned a human-monitoring requirement that indicates whether (and optionally, to what degree) the dialog sequence requires monitoring by a service agent. In addition, instances of the dialog sequences in calls are prioritized based on an indication of caller intelligibility during (i) prior dialog sequences in the current call, and (ii) prior calls involving the same caller and / or the same contact identifier (e.g., telephone number, email address, Internet Protocol address, etc.).